MM-Retinal-Reason
收藏资源简介:
MM-Retinal-Reason 是首个眼科多模态推理数据集,涵盖了从基本推理到复杂推理的完整范围,旨在提升视觉中心的基本推理能力,模拟真实临床思维模式。该数据集由 45 个公共数据集和 PubMed Central (PMC) 收集的实时数据构建,包括 4 种类型的问题,涵盖超过 100 种常见的和罕见的眼科异常。数据集由基本推理视觉问答 (VQA)、思维链 (CoT) 推理轨迹、图像标题对和复杂推理 VQA 四个部分组成,为眼科 MLLMs 的训练和评估提供全面监督。该数据集的创建过程包括数据收集、问题制定、候选选项设计、推理轨迹构建、图像标题对构建和复杂推理 VQA 构建。该数据集的应用领域是眼科诊断,旨在解决真实临床场景中的复杂推理问题。
MM-Retinal-Reason is the first ophthalmic multimodal reasoning dataset covering the full spectrum from basic to complex reasoning. It aims to enhance the fundamental reasoning capabilities of visual-centric systems and simulate real clinical thinking patterns. This dataset is constructed from 45 public datasets and real-world data collected from PubMed Central (PMC), which includes four types of questions covering over 100 common and rare ophthalmic abnormalities. The dataset consists of four components: basic reasoning Visual Question Answering (VQA), Chain-of-Thought (CoT) reasoning traces, image-caption pairs, and complex reasoning VQA, providing comprehensive supervision for the training and evaluation of ophthalmic Multimodal Large Language Models (MLLMs). The construction process of the dataset includes data collection, question formulation, candidate option design, reasoning trace construction, image-caption pair construction, and complex reasoning VQA construction. The dataset is applied in the field of ophthalmic diagnosis, aiming to solve complex reasoning problems in real clinical scenarios.

- 1通过东南大学计算机科学与工程学院, 南京医科大学第一附属医院眼科, 南京理工大学计算机科学与工程学院, 西北工业大学计算机科学学院 · 2025年



